视频图像中的运动目标分割技术是计算机视觉中的一个重要研究方向，是许多视频处理技术的基础。运动对象分割技术具有广泛的应用前景，如视频监控、视频编码、视频检索、多媒体创作、图像理解、模式识别等。对象分割技术的目的是在时间轴上将视频图像分割成一系列相互关联的空间区域。这种基于区域的图像表示方法为视频的灵活处理提供了可能。本文重点研究在摄像机运动情况下的运动目标分割方法，具体的研究内容如下： 一、在全局运动估计方面，提出了一种基于三参数模型的快速全局运动估计算法，主要贡献可以概括为三点：一、提出了新的三参数模型建模全局运动，在保证准确性的同时用更少的参数来描述和估计全局运动；二、提出了基于梯度信息的宏块预判方法，在光流场计算之前提前剔除可能产生错误估计的宏块，通过减少参与光流计算的宏块数目降低参数估计的运算量；三、将梯度信息引入光流场计算中，提出了快速光流场计算算法。 二、提出了一种基于“图切割”（Graph Cuts）算法的运动分层方法：该方法通过定义新的能量函数，使得到的运动层具有更好的连贯性和准确性。此外，为了使检测到的运动目标边缘更加准确，将运动分层结果与静态图像分割结果结合。利用水线法获得图像的静态分割，并通过“水线崩溃”算法进行区域合并改进分割效果，将其与运动分层得到的初始“目标区域”结合，在分割运动目标时得到更加准确的边缘。 三、面向立体视觉在运动目标分割中的应用，提出了一种6参数的立体全局运动模型SGMM（Stereo global motion model）。该模型能够直接利用灰度图像和视差图像的运动变化，估计由摄像机引起的全局运动。基于这一立体全局运动模型，无需知道摄像机的实际运动，就可以快速的区分图像中的全局运动与局部运动，从而为实时的运动目标检测提供了可能。 四、利用立体视觉及SGMM解决交通场景中的运动目标检测问题。首先，通过角点跟踪方法获得连续图像帧之间的运动对应点，根据对应点采用递归最小二乘法估计SGMM的运动参数，由此获得了视频图像中的全局运动信息。然后，通过基于立体坐标系的三维场景分析，将交通场景的图像根据三维特征分割为若干“目标区域”。综合得到的“目标区域”以及全局运动信息，通过将“目标区域”的运动与基于SGMM的全局运动相比较，检测出运动目标。实验结果展现了令人满意的分割效果。
英文摘要
Moving object segmentation in video sequences is a fundamental problem in computervision, and has been studied widely. Moving object segmentation techniques have many applications, such as video surveillance, video coding, video retrieval, multimedia production, image understanding, pattern recognition, etc. The goal of object segmentation is to partition the video frames into a series of spatially correlated regions along the time axis. In this thesis, the research focuses on the problem of moving object segmentation when the video is captured by a moving camera. The detailed description of the above techniques is as follows: 1. A fast algorithm is proposed for global motion estimation, which is based on Iterative Least- Square Estimation (ILSE) technique. A new 3-parameter linear model is proposed to describe and estimate global motion, which is simple and reasonable. Gradient Thresholding (GT) method is introduced to pre-analyze the image macro-blocks before global motion estimation using their gradient information, which reduce the computational cost by reducing the amount of involved blocks. Successive Elimination Algorithm (SEA) is improved to fast calculate motion field. 2. A Graph-cuts based method is proposed to segment moving objects from image sequences. The segmentation framework involves three stages: Motion seed detection, Motion layer expansion and Motion boundary refinement. To preserve the spatial continuity of motion layers, an energy function is defined and Graph-cuts technique is used to solve the energy minimization problem and extract motion layers. 3. Stereovision technique is used to segment moving objects. Since the depth information is available, motion can be estimated more precisely in 3D stereo coordinate system (SCS) constructed by the depth and the image coordinates. A six-parameter global motion model (GMM) is proposed to describe camera motion in SCS. Based on such model, global motion can be estimated and identified directly in SCS without knowing the real physical parameters in WCS about camera motion and camera setup. 4. SGMM is used to detect moving obstacles for ITS (Intelligent Traffic System) applications. Geometric analysis method segments the road scene into 3D object-surfaces based on U-V-disparity features of road surfaces, roadside structures and obstacles. Together with the estimated SGMM parameters, moving surfaces, which correspond to the real moving objects, are extracted.
修改评论